Large language models

Two ways to get a capable LLM into a system: call someone else's model over an API, or download weights and serve them on hardware you control. The trade is capability and convenience against control, data locality and cost at scale.

Checked as of 2026-10-23

The two camps

Frontier

Rent the best model

Closed models from the labs pushing capability: Anthropic, OpenAI, Google. You never see the weights. You send text to an endpoint and pay per million input and output tokens.

The provider handles scaling, upgrades and safety tuning. You get the strongest reasoning, the longest context windows and the broadest tool use, at the cost of sending data to a third party and accepting their release and retirement schedule.

Open-weight

Own a good model

Weights you can download from Hugging Face and run anywhere: Llama, Mistral, Qwen, Gemma, DeepSeek. "Open-weight" is not the same as open source. The licence decides what you may do, and several restrict commercial use or impose acceptable-use terms.

You pay for hardware by the hour whether or not it is busy. In return the data never leaves your account, the model never changes underneath you, and you can fine-tune it.

How to choose

Start with a frontier API	Almost always, for a prototype. It tells you whether an LLM can do the task at all before you invest in hosting. If the best model cannot do it, a smaller open one will not either.
Data cannot leave your boundary	Self-host open weights in your own VPC, or use a frontier model through your cloud provider (Claude on Bedrock, Gemini on Vertex AI) where the provider's data terms are the cloud's, not the lab's.
High, steady volume	A GPU that is busy around the clock can beat per-token pricing. A GPU that idles overnight usually cannot. Run the numbers on the compare page.
Narrow, repetitive task	Classification, extraction, routing: a small open model fine-tuned on your labels is often as accurate as a frontier model, and far cheaper and faster.
Hard reasoning, long documents, agents	Frontier models still lead. The gap narrows every release, so re-test open models periodically.
Reproducibility and audit	Pinned open weights never change. Frontier model IDs are pinned too, but get retired; plan for migration.

Licence classes

Every family on these pages is tagged with one of these. Read the actual licence before shipping; the tag is a summary.

Open source	OSI-style permissive licence (Apache-2.0, MIT). Use, modify, fine-tune and ship commercially with attribution.
Open, restricted	Weights are downloadable and commercial use is allowed, but under the provider's own licence: acceptable-use policies, user-count thresholds or naming requirements.
Non-commercial	Research and personal use only. Production use needs a separate commercial licence from the provider.
Proprietary	No weights. Use through the provider's API or a cloud partner, under their commercial terms.